Both epoll and io_uring depend on the kernel to do the "actual IO", if you want really user-space you need DPDK and a userspace network stack (for TCP/UDP).
Both epoll and io_uring have virtually the same performance. Of course uring is a lot more "ergonomic" that's why it already has amazing momentum.
You can do network IO completion notification, instead of readiness notification with io_uring but not epoll. Which avoids the need for a separate syscall that needs to copy memory into the userspace buffers. That can be noticable.
It should also, at some point, allow for nice zero copy network receive paths under the right circumstances (i.e. the network card DMAing directly into the userspace buffers, without very weird setup/high op overhead).
Both epoll and io_uring have virtually the same performance. Of course uring is a lot more "ergonomic" that's why it already has amazing momentum.